随着计算水平的提升、工业技术的发展，工业系统的功能日益强大，工业设备逐渐朝着复杂化和智能化的方向发展与迭代更新，设备的运行状态和故障情况逐渐趋向非透明化，传统的故障诊断方法不再适用。同时，工业系统的大型化和系统化也对设备的健康管理提出了更高的要求，能否及时、准确地判断故障类型，将影响工业生产的正常进行，关系到人民的财产安全和人身安全。此外，工业设备的智能化使得应用场景拓宽，某种环境和参数下训练的模型在条件改变后将无法适应新的场景。重新采集数据进行标注、训练耗时耗力，在采集及标注过程中还存在难以采集全生命周期的信号、对正在运行的重要设备不可能随时停机进行采集和标注、工人的经验标注不一定准确等等困难。

因此，从海量数据中自动提取特征，训练强大的诊断模型，并适应变化的工作场景具有强烈的现实研究价值。但是，由于深度学习的黑匣子、非透明化、非可解释性等性质，迁移学习在故障诊断领域的广泛部署受到了阻碍，因此可解释性研究十分重要。遵循以上目标，本文设计了基于迁移学习和相似性分析的工业设备故障诊断模型，通过迁移学习帮助模型适应不同的工作环境；通过相似性对迁移学习进行分析和指导，为迁移学习在工业领域的应用提供支撑和保障。主要研究内容与创新如下：

1．针对工业场景中数据无标签的情况，设计了基于子域对齐的工业设备故障诊断模型。该模型将有标签的数据作为源域，将神经网络提取的特征在类别层面上对齐，不同于传统的全局对齐方法，子域对齐可以实现更加细粒度的域对齐，达到更高的诊断准确率。通过在工业数据上开展的多组迁移实验，基于多个评价指标，证明了基于子域对齐的故障诊断模型性能具有优越性；通过观察训练数据量下降情况下模型性能的变化，证明了子域对齐模型在数据量不充足时仍能保持较好的诊断性能。

2．针对由于域差异较大，两个域之间的迁移难度大，导致子域对齐效果不佳的现象，本文将子域对齐模型给目标域分配的标签视为有错误的标签，将其泛化为一个弱监督问题，使用带噪标签学习的思想进一步提升诊断准确率。本文为每一个数据创建一个标签分布，然后在每次迭代中同时更新网络参数和分布，概率地纠正错误标签。通过实验验证，本文提出的模型可以在子域对齐结果的基础上纠正噪声标签，进一步提升诊断准确率，并且整个训练过程都不需要目标域标签数据的参与。

3．针对深度学习的黑匣子特性，本文从相似性的角度切入，借助相似性指标对深度学习、迁移学习、带噪标签学习过程进行分析和解释，并进一步对迁移学习的源域选择、迁移层的选择等问题进行指导。本文基于中心核对齐、表示相似性分析、对数期望经验预测等指标，对具体的训练过程进行了分析、解释和指导，通过实验验证了指标的合理性。

With the development of computing power and sicence, industrial systems are becoming increasingly complex, and the equipment is gradually developing in the direction of large-scale, high-speed, systematization and automation. The health condition of the equipment has gradually become non-transparent, and the traditional fault diagnosis methods are no longer applicable. At the same time, the large-scale and systematic industrial system also puts forward higher requirements for the health management, whether the fault type can be judged timely and accurately will affect the normal industrial production and people's property safety. In addition, the intellectualization of industrial equipment has broadened the application scenarios, and the model trained under the specific environment will not be able to adapt to new scenarios after the conditions change. Re-collecting data for labeling and training is time-consuming and labor-intensive. Besides, it is very difficult to collect data throughout the whole life cycle, and the labels annotated by engineers based on their experience may not be accurate.

Therefore, it is critical to automatically extract features from massive data, train powerful diagnostic models, and adapt to changing work scenarios. However, the black box, non-transparency, and non-interpretability nature of deep learning have hindered the widespread deployment of transfer learning in fault diagnosis. Therefore, the research on interpretability is very important. According to the above problems, this paper presented a fault diagnosis model for industrial equipment based on transfer learning and similarity analysis. Transfer learning helps the model adapt to different working environments and similarity analysis supports the development of transfer learning in fault diagnosis. The main research contents include the following three parts:

1. A fault diagnosis model for industrial equipment based on subdomain alignment is proposed for the data without labels. In this model, the features of the source and target domain extracted by neural network are aligned at the category level. Different from the traditional global alignment methods, subdomain alignment can achieve fine-grained domain adaptation, and the performance of the proposed model is proved to be superior through diverse transfer experiments and several evaluation indicators designed in this paper. In addition, the presented model can still maintain good performance when the amount of training data decreases.

2. Transfer learning is more difficult when the shift between the source and target domain is large, and resulting in poor performance. In this paper, the labels assigned to the target domain by the subdomain alignment model are regarded as noisy labels, which is generalized as a weak supervision problem, and the learning with noisy labels is used to further improve the diagnostic performance. Specifically, we create a label distribution for each data, and update the network parameters and the distributions simultaneously by the designed loss function in each iteration to correct labels. Experimental results show that the proposed model can correct noisy labels introduced by the misalignment of the subdomain adaptation, further improve the diagnostic accuracy, and the whole training process does not require labeled data of the target domain.

3. With the help of similarity index, this paper analyzes the process of deep learning, transfer learning and learning with noisy labels to explaine the phenomenon in the training and further guides the selection of the source domain and the adaptive layer. In this paper, the Centered Kernel Alignment, Representation Similarity Analysis, Log Expected Empirical Prediction and Earth Mover’s Distance are selected, and the effectiveness and reliability of these indicators are verified by diverse experiments.